acceptodds
Under review as a conference paper at ICLR 2027

Dual-Branch Chinese Spelling Correction Method for Target Personal Names

Abstract

Chinese personal name correction is an important problem in Chinese spelling correction (CSC) and has been widely applied in search correction, writing assistance, and post-processing of recognition results. Existing name correction methods often rely too heavily on name boundary detection, candidate name sets, or external knowledge resources, making them susceptible to biases caused by errors in obtaining candidate names and insufficient resource coverage. To address these problems, we incorporate Chinese personal name correction into a unified CSC framework and propose a Dual-Branch Chinese Spelling Correction (DCSC) method for the character-level correction of target personal names. Firstly, an online correction dataset for Chinese target person names is constructed. This dataset was created by replacing the personal pronouns in the existing Chinese corpus and generating a series of erroneous name samples, thereby avoiding the problem of incorrect candidate names due to errors in name boundary detection. Furthermore, we present a dual-branch gated fusion module. This module consists of a name-specific branch and a general branch, which respectively model the name-specific features and the general correction features. Through this module, Chinese personal name correction and general CSC are integrated into a single framework. In addition, a gating mechanism is employed to automatically fuse the features obtained by the two branches, thereby reducing the interference of the training for specific names on the overall correction ability of the model. DCSC not only can accurately correct errors in personal names, but also does not significantly reduce the overall spelling correction performance of the model. On the three classical CSC models, namely, MDCSpell, SCOPE, and NamBert, our method achieved F1 scores of 94.14%, 95.10%, and 93.95%, respectively, on the proposed online Chinese personal name correction dataset. This indicates that our method can effectively correct both Chinese personal names and common spelling errors simultaneously.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Chinese spelling correction, Chinese personal name correction, dual-branch architecture

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_AdrCOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper adapts Chinese spelling correctors to a supplied list of personal names. It builds training examples by replacing pronouns and introducing phonetic or glyph-based errors, then adds dual-branch gated fusion (DBGF) to help preserve general spelling-correction ability.

Critical strengths

The task has a clear practical motivation: a misspelled name can still look plausible, making it difficult to correct from context alone. Pronoun replacement supplies contextual examples while retaining ordinary spelling errors. The gains on CPN-Corr are large: correction F1 rises from 38.45 to 89.54 for MDCSpell, 37.72 to 94.33 for SCOPE, and 35.40 to 91.47 for NamBert (Table 1). The three backbones provide useful evidence that the recipe transfers between correctors. Figures 1–3 explain the motivation and construction clearly.

Critical weaknesses

My main concern is what the paper adds beyond the existing adaptation recipe. CN121435969A, published 2026-01-30, already describes pronoun replacement, three confusion types, and CSC adaptation to supplied names. That leaves DBGF and the additional experiments as the main additions. Section 3.3, however, gives both branches the same input and type of transformation without explaining how one learns name correction and the other general correction. Table 2 shows that the module helps retain general ability, but does not distinguish the proposed division of roles from simply adding comparable residual capacity. Its CPN-Corr effect is -0.08 to +1.32 points; most of the name-correction gain comes from data adaptation.

The data splits are also difficult to assess. Sections 3.2 and 4.1 give the final sizes, but not the original SIGHAN/Wang271K partitions or whether variants of a source sentence stay in the same split. I cannot tell whether the evaluation uses independent contexts or whether adaptation data overlaps with the reported SIGHAN test sets.

Questions for authors

Which original partitions were used to build each CPN-Corr split? Were variants grouped by source sentence, and were any SIGHAN test sentences used for adaptation?

Does Section 3.3 omit any branch-specific objectives, initialization choices, or update schedules? Is there evidence for the assigned roles beyond the ablation of the whole module?

1 comment

Critical weakness My main concern is what the paper adds beyond the existing adaptation recipe. CN121435969A (https://patents.google.com/patent/CN121435969A/zh), published 2026-01-30, already describes pronoun replacement, three confusion types, and CSC adaptation to supplied names. That leaves DBGF and the additional experiments as the main additions. Section 3.3, however, gives both branches the same input and type of transformation without explaining how one learns name correction and the other general correction. Table 2 shows that the module helps retain general ability, but does not distinguish the proposed division of roles from simply adding comparable residual capacity. Its CPN-Corr effect is -0.08 to +1.32 points; most of the name-correction gain comes from data adaptation. Authors’ response: Thanks very much for your valuable comment. CN121435969A is a patent that reports previous work of our team. Building on that method, this paper introduces a dual-branch gated fusion module, together with a controlled evaluation of the module that keeps the training data fixed and varies only the module. The improvement on CPN-Corr comes from the data rather than from the module, and we say so rather than leave it to inference. Questions for Authors

  1. Which original partitions were used to build each CPN-Corr split? Were variants grouped by source sentence, and were any SIGHAN test sentences used for adaptation? Authors’ response: Thanks very much for your valuable comment. The source pool, and the way in which the split is produced, are as follows. All three CPN-Corr splits were built from the same source pool, namely the training partitions of SIGHAN13, SIGHAN14, SIGHAN15 and Wang271K. There is no one-to-one mapping between an original partition and a CPN-Corr split: sentences from the four training partitions were pooled, the name-targeted samples were generated from the pooled sentences, and the 15,000 generated samples were then randomly divided into training, validation and test sets in the proportions of 50%, 33% and 17%. On the grouping of variants: the division is random at the level of the generated samples, so the variants of a source sentence are not kept together, and we do not claim that they are. On the SIGHAN test material: no test sentence was used for adaptation or for any other part of the construction. The exclusion is decided over the source sentences, before replacement and perturbation.
  2. Does Section 3.3 omit any branch-specific objectives, initialization choices, or update schedules? Is there evidence for the assigned roles beyond the ablation of the whole module? Authors’ response: Thanks very much for your valuable comment. What the branches share, what differs between them, and the evidence for the assigned roles are as follows. No branch-specific objective, initialization or update schedule is omitted, because there are none. Both branches receive the same input, have the same architecture, a two-layer MLP, and are trained end-to-end under a single correction loss. No auxiliary loss, no contrastive or orthogonality term and no branch-specific label is used, and all parameters are updated jointly at every step, with nothing frozen or alternated. The absence of branch-specific supervision is stated rather than left to inference. The only per-branch difference that exists is separate learning rates for the encoder, the gate, the name branch and the general branch, together with the initialization of the gate bias (1, −1 and −1.5 for NamBert, SCOPE and MDCSpell) and of the residual scale α (0.2, 0.2 and 0.5). These were chosen on the validation split of CPN-Corr, and they did not produce a consistent pattern across the three backbones. Beyond the whole-module ablation there is no evidence for the assigned roles, and we do not claim any. The ablation cannot separate the two-branch structure from the added parameters, and the current evidence cannot establish functionally distinct representations. The per-group learning rates and initializations were our attempt to obtain such evidence, since a branch that receives a larger learning rate, or a gate initialized towards it, should absorb more of the name-specific signal. The attempt produced no clear pattern and is reported as a negative result.
Reviewer_SyN5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the Chinese spelling correction problem for target personal names and proposes a dual-branch Chinese spelling correction framework, namely DCSC. The method first constructs a target-name correction dataset by replacing personal pronouns in existing CSC corpora and generating homophonic, near-phonetic, and orthographic errors. It then introduces a dual-branch gated fusion module to integrate name-specific and general correction features, aiming to improve personal-name correction while maintaining general CSC performance. The paper also conducts experiments on three CSC models and multiple datasets to demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The proposed data construction strategy provides a relatively, simple and practical way to generate training samples for specific target names, which can help alleviate the lack of name-related correction data in existing CSC datasets.
  2. The dual-branch gated fusion mechanism is a reasonable modeling design. It attempts to balance name-specific correction and general spelling correction, while avoiding explicit name boundary detection and dependence on external candidate-name resources.
Critical weaknesses

1.The constructed CPN-Corr dataset contains only five target personal names, with 3,000 samples for each name. Moreover, the dataset is generated by replacing personal pronouns in existing CSC corpora and introducing synthetic spelling errors. Such a construction process may not adequately reflect the distribution and complexity of personal-name errors in real-world scenarios. The strong performance on this dataset may mainly reflect the model's adaptation to a small number of predefined target names and their synthetic error patterns.

2.The motivation behind the dual-branch architecture is reasonable, but the actual functional distinction between the two branches is not sufficiently justified. As shown in Equations (3)–(6), both branches receive the same backbone representations and are combined through a learnable gating mechanism. However, the paper does not introduce explicit supervision or constraints to ensure that one branch learns name-specific features while the other preserves general correction knowledge. Therefore, it remains unclear whether the two branches truly learn complementary representations or simply introduce additional model capacity. The paper needs to provide more convincing explanations or analyses of the branch specialization mechanism.

Questions for authors

Q1: The data splitting procedure of CPN-Corr is not clearly described. It is unclear whether the original sentences were divided into training, validation, and test sets before generating different name-error samples, or whether the generated samples were randomly split afterward. Q2: Since the CPN-Corr dataset is constructed using existing CSC corpora, including SIGHAN, which is also used for general CSC evaluation, it is important to clarify whether the official SIGHAN test sentences were excluded from the construction of CPN-Corr training data. Otherwise, potential data overlap could affect the reliability of the evaluation results.

Q3: The CPN-Corr dataset contains only five target personal names, and all name-related errors are synthetically generated. It remains unclear whether the proposed method can generalize to a larger number of target names and naturally occurring spelling errors.

Q4: The functional specialization between the two branches needs further justification. Since both branches receive the same backbone representations and are trained within the same framework, what mechanism encourages the name-specific branch to learn personal-name features while the general branch preserves general correction knowledge? Q5: The ablation studies are not sufficiently targeted to demonstrate the necessity of the proposed dual-branch gated fusion design. Although Table 2 compares the models with and without DBGF, the contributions of the dual-branch structure, adaptive gating, and additional model parameters are not separately examined. Comparisons with a parameter-matched single-branch residual module and a dual-branch model using fixed fusion weights are needed to better justify the proposed design.

Q6: The experiments lack direct comparisons with existing personal-name correction methods. Such comparisons are needed to demonstrate the effectiveness and advantages of the proposed method.

Q7: How does the proposed framework handle newly added target personal names? Is additional fine-tuning required for each new name?

Reviewer_fqPAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies correction of pre-specified Chinese personal names within a character-level Chinese spelling correction (CSC) framework. It constructs CPN-Corr by replacing personal pronouns in existing CSC corpora with target names and perturbing the names with homophonic, near-phonetic, and visually similar characters. It also proposes a Dual-Branch Gated Fusion (DBGF) module, which is intended to combine name-specific and general correction features. Experiments use MDCSpell, SCOPE, and NamBert as backbones. The results show large gains on the authors’ synthetic name dataset, but all nine enhanced-versus-base comparisons on SIGHAN show lower correction F1. The evidence is limited by the five-name closed-set evaluation, insufficient dataset-lineage and reproducibility details, and the lack of experiments establishing that the two DBGF branches have the claimed distinct roles.

Critical strengths

The paper addresses a practical problem: low-frequency personal names can be difficult for general CSC systems to correct from context, especially when the misspelling is itself a plausible word. Framing correction of pre-specified names as character-level CSC is a clear and potentially useful direction.

The data-construction idea is straightforward, and the paper evaluates three different CSC backbones rather than only one. It also includes a w/o-DBGF comparison. I recalculated the reported F1 values from precision and recall in Tables 1 and 2; they are consistent within rounding error.

Critical weaknesses

The core name-correction evaluation is narrow. CPN-Corr contains only five target names, with synthetic errors generated using the paper’s own homophonic, near-phonetic, and glyph-similarity procedures. The paper does not list the names or establish that test names, characters, or confusion pairs are held out from training. Thus, the results demonstrate performance on a small closed set under a matching synthetic process, but do not establish generalization to unseen names or realistic name errors.

The relationship between CPN-Corr and the SIGHAN evaluation sets is unclear. Section 3.2 says that CPN-Corr is reconstructed and expanded from SIGHAN13/14/15 and Wang271K, and that sentences containing pronouns are selected from these corpora. Section 4.1 then evaluates on SIGHAN13/14/15. The paper does not describe source-split lineage, whether splitting occurred before augmentation, or whether SIGHAN test sentences and near-duplicates were excluded. This does not prove leakage, but leaves a material train–test contamination risk unresolved. The authors should provide source-level lineage and demonstrate zero overlap with the SIGHAN test sets.

The main comparison conflates name-oriented fine-tuning and DBGF: the “Ours” models receive both, whereas the base models receive neither. Table 2 partially isolates DBGF, but only on SIGHAN15 and CPN-Corr. On CPN-Corr, DBGF changes correction F1 by −0.08 for MDCSpell, +0.65 for SCOPE, and +1.32 for NamBert. The large gains in Table 1 therefore cannot be attributed to DBGF. Moreover, both branches are defined as functions of the same backbone representation, with no branch-specific supervision or constraint to make the “name” and “general” roles identifiable.

The claim that general CSC is not significantly degraded is inconsistent with Table 1 and unsupported statistically. All nine enhanced-versus-base SIGHAN correction-F1 comparisons are lower. The mean change is −1.334 percentage points, with a range from −2.81 to −0.50. Detection F1 also decreases in all nine comparisons, by −1.134 points on average. No multiple-seed results, confidence intervals, or significance tests are reported. The abstract also describes 94.14%, 95.10%, and 93.95% as unqualified F1 scores, although these are detection F1; the corresponding correction F1 scores are 89.54%, 94.33%, and 91.47%.

The method is not sufficiently specified for reproduction. Important details are missing for the confusion sets (including the top-K value for glyph candidates), source-data filtering and split construction, model architecture, training protocol, random seeds, and evaluation definitions. The paper also lacks the AI-use statement required by the ICLR 2027 author guidelines. The English prose and figure text need careful editing; for example, Figure 3 translates “我喜欢吃苹果” as “I like eating flat fruit.” These editing issues are not evidence of AI use or misconduct, but the required disclosure section is a concrete compliance issue.

Questions for authors

1.How were the source sentences assigned to the CPN-Corr train/validation/test splits? Was splitting performed before name replacement and perturbation, and can you verify that CPN-Corr training data contain no SIGHAN test sentences or near-duplicates?

2.Since CPN-Corr contains only five target names and synthetic errors, do the train and test sets share names or confusion patterns? Can you provide an evaluation on unseen names or an independent set of realistic name errors, or clarify that the method is limited to adapting to pre-specified names?

3.Can you separate the effect of CPN-Corr fine-tuning from the effect of DBGF with matched-budget comparisons? Given that both branches receive the same representation, what evidence supports their distinct “name” and “general” roles?

4.All nine enhanced-versus-base SIGHAN correction-F1 comparisons are lower, yet the paper claims no significant degradation. Can you report results across multiple seeds with uncertainty estimates, clarify the evaluation protocol, and revise this claim accordingly?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.